Neural Episodic Control

Abstract

Deep reinforcement learning methods attain super-human performance in a wide range of environments. Such methods are grossly inefficient, often taking orders of magnitudes more data than humans to achieve reasonable performance. We propose Neural Episodic Control: a deep reinforcement learning agent that is able to rapidly assimilate new experiences and act upon them. Our agent uses a semi-tabular representation of the value function: a buffer of past experience containing slowly changing state representations and rapidly updated estimates of the value function. We show across a wide range of environments that our agent learns significantly faster than other state-of-the-art, general purpose deep reinforcement learning agents.

References in corpus (4)

• PathNet: Evolution Channels Gradient Descent in Super Neural Networks
• Learning to Remember Rare Events
• Hybrid Reward Architecture for Reinforcement Learning
• Learning to Play in a Day: Faster Deep Reinforcement Learning by Optimality Tightening

Cited by in corpus (10)

• Stabilizing Transformers for Reinforcement Learning
• A Short Survey On Memory Based Reinforcement Learning
• Concurrent Meta Reinforcement Learning
• Biologically inspired architectures for sample-efficient deep reinforcement learning
• Memory-Efficient Episodic Control Reinforcement Learning with Dynamic Online k-means
• Sample-Efficient Reinforcement Learning with Maximum Entropy Mellowmax Episodic Control
• Continual and Multi-task Reinforcement Learning With Shared Episodic Memory
• Random Projection in Neural Episodic Control
• Asynchronous Episodic Deep Deterministic Policy Gradient: Towards Continuous Control in Computationally Complex Environments
• Improving Sample Efficiency with Normalized RBF Kernels